Conversation with Yidian Tianxia: Agent Has Essential Differences from the Past, Security Risks Lie Not in Model Fine-tuning, But in Model Selection

Invest in stocks with the Golden Kylin Analyst Reports—authoritative, professional, timely, comprehensive—helping you uncover potential thematic opportunities!

(Source: Titanium Media APP)

In recent years, generative AI technology has advanced rapidly. For outbound companies, the biggest challenge is not higher traffic costs, but a change in the underlying logic of growth. A real issue is that the outbound marketing industry is undergoing a profound shift from a labor-intensive model to an AI-driven one.

As a marketing technology company, Yidian Tianxia provides two core services for outbound businesses: first, global brand expansion, offering a one-stop overseas marketing solution from exposure to conversion; second, leveraging algorithms and data to enable advertisers to achieve intelligent ad placement.

With the exponential increase in global complexity, brand customer acquisition costs continue to rise, and the return on targeted ad spend declines year by year. Traditional methods of increasing budgets or adding manpower are no longer effective. Additionally, intelligent decision-making has experienced a disconnect—despite platforms transmitting large amounts of data daily, the volume of data cannot be converted into real-time decisions.

Yidian Tianxia’s Chief Algorithm Scientist, Ady Zhao, told me that the development and maturity of Agent technology is fundamentally different from the past. Previously, Agents were question-and-answer models; now, Agents can truly perceive, autonomously plan, make decisions, and execute, achieving end-to-end closed loops. Agents are no longer just auxiliary and execution tools—they are beginning to take over the entire chain from insights to execution. In other words, humans are shifting from operators to decision-makers.

It is reported that Yidian Tianxia has been laying out AI since the GPT era, building a complete marketing product matrix and solutions around the entire marketing chain, enabling intelligent insights, creative generation, automated delivery, and data attribution—forming a full-chain marketing AI automation loop.

From large models to Agents, circulating through an AI middle platform

AI large model technology is gradually penetrating various industries, and its implementation effects and replicability have always attracted attention.

Yidian Tianxia’s Chief Product Officer, Aodi Zhang, pointed out that the biggest challenge for AI in marketing and advertising is that content generated by large models does not necessarily guarantee results. In other words, AI can quickly generate ad materials, but it cannot ensure effective delivery.

To address this, Yidian Tianxia has focused on two aspects: first, a systematic approach that combines the existing data middle platform to quickly review results, select high-quality materials, and enable large-scale production and deployment; second, for vertical scenarios, through fine-tuning or orchestration, building vertical models—such as one-click generation of effective materials.

“Because current models are point-based and not integrated with business, during marketing, we still need to provide decision Agents with decay data like final ROI or customer lifetime value (LTV) for each generated content, to judge quality, adjust strategies, and allocate budgets. This is a systematic project—an integration of traditional marketing and AI,” said Aodi Zhang.

This led Yidian Tianxia to build an enterprise-level AI agent development platform, EC-Agent, which can access mainstream large model providers’ models in one go. Through evaluation, it determines which AI models are suitable for different stages of marketing. Meanwhile, on the front end, Agents are set up to schedule different tasks, coordinated via EC-Agent, which manages various tools. The approach involves constructing multiple Agents based on scenarios and business units, simulating roles like product managers, R&D, and designers. By standardizing actions and evaluating reusable versus personalized capabilities—such as experience-based personalized content—based on business units, Agents are integrated into EC-Agent. Currently, EC-Agent operates over 200 Agents. Building on this, an application layer is developed, continuously accumulating niche models and small Agents, encapsulating them into products.

It is reported that Yidian Tianxia mainly built the EC-Agent using Amazon Web Services’ Agentic AI technology, including services like Amazon Bedrock AgentCore, open-source Agent development framework Strands Agents, Amazon Bedrock knowledge bases, Amazon Nova models, and more. These support internal business needs and innovative explorations, enhancing internal efficiency and accelerating AI application deployment.

In business operations, EC-Agent manages the entire ad cycle—from market research and audience profiling to creative generation and intelligent delivery strategy optimization—deeply optimizing the ad chain. Data shows that, with this platform, clients can reduce ad launch time from 5 days to 2 hours, with automation reaching 80%.

In terms of intelligent delivery, previously, budget allocation, bid adjustments, and creative changes relied on the experience of optimization specialists. Now, with the AI-driven intelligent delivery system, humans shift from operators to decision-makers and strategists. An optimizer who managed 30 campaigns before can now handle 300 or more, with better results. Currently, clients can reduce 15-20% of ad waste through AI optimization, with decision response times shortened from hours to minutes.

For internal efficiency, EC-Agent not only provides dedicated Agents for operations, maintenance, and business departments but also offers an OA Agent for all employees, allowing staff to focus on more creative business scenarios and driving organizational efficiency to new heights.

If EC-Agent is the “factory that builds cars,” then AI Drive 2.0 is the “finished vehicle” delivered to clients. Currently, Yidian Tianxia has implemented the new generation of intelligent marketing solutions, AI Drive 2.0, based on EC-Agent, and has built an AI product matrix including FunsData, KreadoAI, CyberGrow, AdsGo.ai, Cycor, among others, to coordinate multiple Agent capabilities.

Defining boundaries for Agent authority

As Agent capabilities continue to deepen in the enterprise market—especially with the recent surge of local deployment tools like OpenClaw—security risks become more prominent.

For example, risks such as AI hallucinations, prompt injection attacks, and open-source skills tools used in AI scheduling could be amplified in Agent applications. Autonomous decision-making by Agents also introduces risks—whether exploited by malicious actors or due to their inherent completeness and capabilities—leading to uncontrollable and unpredictable risks.

Aodi Zhang pointed out, “I believe the biggest risk isn’t model fine-tuning, but model selection. We classify data—for example, handling sensitive data like financial or customer contracts—using private deployment models. In Agent scheduling, the core is what tools we provide.”

To this end, Yidian Tianxia’s EC-Agent, including tools like Skills, employs whitelist mechanisms, requiring each tool to be reviewed by IT and security teams. Even if an Agent makes an erroneous call, it remains within the service scope. The accuracy of calls is based either on SOP workflows or the Agent’s own intelligence, ultimately measured by business outcomes.

Ady Zhao emphasized that for core business scenarios to be effectively implemented, control is key. Their approach involves three steps:

In Aodi Zhang’s view, technological breakthroughs have opened new horizons, but truly integrating AI into core business relies on a comprehensive engineering approach combining “model capability + business knowledge + tool constraints + data fine-tuning.” Having a good model alone isn’t enough; it must be used correctly, reliably, and precisely.